1
Whose Move is it Anyway? Authenticating Smart
Wearable Devices Using Unique Head Movement
Patterns
Sugang Li∗ , Ashwin Ashok† , Yanyong Zhang∗ , Chenren Xu‡ , Janne Lindqvist∗ , Macro Gruteser∗
∗ WINLAB, Rutgers University, North Brunswick, NJ, USA
† Carnegie Mellon University, Pittsburgh, PA, USA
‡ CECA, Peking University, Beijing, China
Abstract—In this paper, we present the design, implementation
and evaluation of a user authentication system, Headbanger,
for smart head-worn devices, through monitoring the user’s
unique head-movement patterns in response to an external audio
stimulus. Compared to today’s solutions, which primarily rely on
indirect authentication mechanisms via the user’s smartphone,
thus cumbersome and susceptible to adversary intrusions, the
proposed head-movement based authentication provides an accurate, robust, light-weight and convenient solution.
Through extensive experimental evaluation with 95 participants, we show that our mechanism can accurately authenticate
users with an average true acceptance rate of 95.57% while
keeping the average false acceptance rate of 4.43%. We also
show that even simple head-movement patterns are robust against
imitation attacks. Finally, we demonstrate our authentication
algorithm is rather light-weight: the overall processing latency
on Google Glass is around 1.9 seconds.
I. I NTRODUCTION
Wearable devices are on the way to become an integral
part of people’s daily lives [10], [16], [35]. These devices
collect data about the wearers, their surroundings and often
even about their health. It is thus critical to the users’ privacy,
that this data is protected from unauthorized access. Although
there has been work [19], [20], [23] on limiting privacy threats
from ubiquitous photography enabled by the wearable devices,
robust usable and secure authentication systems leveraging
the devices have not emerged. An authentication system for
these devices has to strike an appropriate balance with user
convenience, especially since users are interacting with an
increasing number of wearables.
Authentication Challenge. Today, authentication on most
commercially available wearable devices [10], [35] relies on an
indirect mechanism, where users can log in to their wearables
through phones. This requires the wearable device to be registered and paired to the mobile device, and both devices to be
carried by the user, which can be highly inconvenient in reality.
The security of this approach is also in question as it increases
the chance of hacking into both devices if either is lost or
stolen. Some devices including Google Glass [16] and FitBit’s
health tracker [10] allow linking the device to online accounts
instead of a mobile device for convenience, which, however,
does not make it more secure. Indirect authentication remains
Fig. 1. Illustration of Headbanger. The head-worn device authenticates the
users based on signatures generated from head-movement patterns. These
patterns are created in response to an audio snapshot played on the device.
a dominant paradigm for wearables despite these fundamental
shortcomings because such devices are seriously resourceconstrained in many aspects: battery power, computational and
storage capabilities, and input/output. As a result, typical authentication methods designed for more resource-rich devices
can not be directly applied here; rather, user authentication
for wearable devices must operate indirectly through a more
capable device. We, however, take the viewpoint that wearables
will become more independent units that have to maintain
security guarantees without such paired devices and we seek
to develop suitable direct authentication methods that are both
accurate and light-weight.
Before we explore direct authentication methods for wearable devices, let us first consider available solutions for other
mobile systems, especially smartphones and tablets. Broadly
speaking, the two most commonly used authentication methods
on mobile systems are (arguably) password-based methods
(with their variants) and biometric-based methods. However,
we argue that neither of these two methods is really suitable for
wearable devices. Typing passwords or drawing swipe patterns
on wearable devices can be rather cumbersome due to their
small input/output units, if they do have a touch sensor at all.
Collecting and recognizing physiological biometrics (such as
DNA, fingerprint, hand/finger geometry, iris, odor, palm-print,
retinal scan, voice) requires specialized sensing hardware and
processing resources that add cost, and many of these sensors
are even larger than the size of wearables themselves.
We therefore focus on a third class of direct authentication
methods: relying upon the uniqueness of human behavior
characteristics such as human walking gait, arm swings, typing
2
patterns, body pulse beats, eye-blinks, etc. This way of authenticating users is often referred to as behavior-based authentication, and it has been studied in the context of authenticating
smartphones and tablets [6]–[8], [25], [28]–[30], [36]. The
main advantage of using behavioral characteristics for mobile
devices is that the signatures can be readily generated from
raw data of built in sensors such as motion sensors, camera,
and microphones. Considering that cameras and microphones,
as well as vision and audio processing algorithms, are quite
energy-hungry, we thus focus on those behavioral characteristics that can be easily captured by sensors that require less
power consumption, such as accelerometer. More specifically,
we propose to authenticate wearable devices to users based
on the following behavioral characteristic: our unique body
movement patterns and their dependence on external stimuli
that wearable devices can generate, such as vibrations and
music.
Head-movement based authentication. Body movement patterns have long been used by humans to discriminate between
people. By watching how a person walks, dances, waves
hands, we can often recognize the person from afar. This is
because human body movements are usually distinctive and
repeatable. Achieving the same through wearables, however, is
not straightforward and poses significant research challenges:
it is unclear whether these seriously-constrained devices are
able to capture the differentiating characteristics of movement
patterns, process the data, and quantify the uniqueness of
each user’s behaviors. Moreover, each device will have only a
limited view of body movements, dependent on its mounting
position on the human body. In this paper, we set out to
conduct a holistic study of wearable authentication through
body movements and to design an accurate, robust and lightweight authentication system. A key distinguishing feature of
our work is that we will also consider stimuli that wearable
devices can provide, particularly stimuli that are difficult to
observe even for the closest adversaries. For example, we can
use fast-tempo music through earbuds to stimulate movements
and to make such free-style movements more repeatable.
In particular, we have designed, implemented and evaluated
Headbanger, an authentication system that can authenticate
users by sensing head movements when listening to music
beats. Although we use Google Glass as a running example,
our design can be applied to other head-worn gadgets and
any system that can record head-movements through motion
sensing. Our choice for using head movements is motivated by
the fact that head-worn wearables are becoming very common
today and such devices are already equipped with motion
sensors; for example, personal imaging and heads-up display
devices, gaming headsets, augmented reality devices.
In summary, the key contributions of this paper are:
1) We have designed and implemented a novel user authentication method for wearable devices using head-movement
patterns. Our study shows that user’s head-movement
patterns contain unique signatures that when inferred
correctly can be used as valid means for authentication.
We design a system, Headbanger, that records, processes,
and classifies head-movement patterns of users based on
the built-in accelerometer sensor readings.
2) Through comprehensive experiments involving 95 participants and over different system design parameters we
show that head-movement patterns can generate accurate
authentication results. Our approach effectively identifies
a wearable device user, with an average false acceptance
rate of 4.43% and an average true-positive rate of 95.57%.
Also, we show that our authentication method is quite
robust: when a user slightly increases her head-movement
complexity, it quickly becomes much harder for attackers
to imitate the movement pattern.
3) We implement Headbanger on Google Glass and carefully profile the execution time of each software module
in the implementation. Our measurements indicate an
average processing latency of 1.93 seconds on the Google
Glass for the best authentication performance.
II.
BACKGROUND
A. Mobile Device Authentication Through Body Movements
A number of body-movement based authentication approaches have been proposed for mobile devices. These systems leverage unique signatures from human behavior that may
be subconscious or in response to external stimulus or both.
For example, it has been shown that gait (e.g. stride length,
the amount of arm swing) when the user is walking or running
is a reliable identification cue, and irrespective of the environment [36]. Okumura et. al. [29] have shown that human arm
swing patterns can be used to create signatures to authenticate
to their cell-phones. Monrose et.al. [28] show that keystroke
rhythms, when users type on the keyboard, that include typing
dynamics such as how long is a keystroke, how far is between
consecutive strokes, and how is the pressure exerted on each
key, can be used to authenticate users. Similarly, mouse usage
dynamics [25] and touchpad touching dynamics [3], [7] have
also been shown to serve as potential authentication cues.
We take the viewpoint that, in comparison to other means
of authentication, body-movement based authentication may
offer great convenience. With this rationale, we design an
authentication system, dubbed Headbanger, for head-worn
devices by monitoring user’s unique head-movement patterns
in response to an external audio stimulus.
B. Using Head-movement for Authentication
According to Jain et al. [22], a type of body movement
is useful for authentication when it is universal, distinctive, repeatable, and collectible. Sensors for collecting headmovement patterns are available on most of today’s headworn wearable devices, and thus making head movements both
universal and collectible.
In this paper, we show that free-style head movements are
distinctive and repeatable, especially when combined with
external stimuli such as music beats. In Headbanger, music
plays a crucial role in stimulating body movements such that
the resulting movement pattern is natural to the user (more
distinctive) and easier to remember (more repeatable). Zentner
and Eerola [38] have shown that most people move their body
3
Response
Time
Motion
Peak
Beat
Fig. 2. The response time of a head motion is the interval between the motion
and the music beat to which the motion responds. From a sequence of head
motions, we can obtain the response time sequence.
as a natural response to external rhythmic stimuli such as
music; even at a very early age, infants respond to music
and their movements speed up with the increasing rhythm
speed. Most adults naturally perform head movements or hand
movements when listening to a fast beat audio track [24].
When combined with external rhythmic stimuli, we believe
body movements become more distinctive – not only a person’s
movement pattern is unique, but their response to rhythmic
stimuli is also unique. In this way, the resulting authentication
system will be more dependable.
C. Motivation for Headbanger
Next, we conducted a preliminary experiment to investigate
whether head-movement patterns can be potentially used to
authenticate users. In this experiment, we collected headmovement accelerometer data from 28 subjects, wherein each
subject was asked to perform a simple nodding movement
following a short audio track (referred to as music cue in the
rest of the paper). For this purpose, we examined a simple
aspect of the head-movement pattern, response time, which is
the time interval between a music beat and the corresponding
nodding motion, as shown in Figure 2. The response time
indicates how quickly a person responds to music beats.
In this experiment, we collected 20 samples from each
subject, hence 20 response time time series for each subject.
Then we calculated the average distance scores between these
response times, both for the same subject and across subjects.
We considered three types of distance metrics: cosine distance
(COS), correlation distance (COR), and dynamic time warping
(DTW) distance [2], and the readers can find the detailed
description of these three metrics in Section III-B. We plot
these three types of distance values in Figures 3 (a)-(c),
respectively. In each plot, we include distance values for all
28 subjects – for each subject, we plot average distance scores
between her and each of the other 27 subjects (referred to
as distances with false subjects, shown in blue dots), as well
as average distance scores among her own response times
(referred to as distances with true subjects, shown in red
squares). All three figures clearly demonstrate that the average
Fig. 4. Headbanger system design flow. The online authentication phase of
Headbanger consists of the following steps: (1) sensor data collection in which
we collect accelerometer data while users move their head as a response to an
audio track played on the glass, (2) filtering in which we apply a Butterworth
filtering to smoothen the sensor data for subsequent processing, (3) parameter
generation in which we calculate the distances between two accelerometer
samples as the parameter, and (4) classification in which we adopt an adaptive
thresholding mechanism to classify the user’s head movement, whose output
will be used as the authentication result.
distance score between a subject’s samples is much lower than
that among different subjects’ samples, which further suggests
that a subject exhibits repeatable and unique head nodding
characteristics.
These observations suggest that even with simple nodding
movements, the accelerometer data collected by Google Glass
have the potential to be used for accurate and robust authentication. Motivated by this observation, we next devise the
Headbanger authentication system.
III.
H EADBANGER S YSTEM D ESIGN
Headbanger enables direct authentication of users to
their smart-glass devices or smart-glass apps using headmovements. We posit that Headbanger will run as a service
in the device upon power-up or application start-up, similar to
the screen-lock in smartphones.
The authentication process has two phases: an offline training phase and an online authentication phase. In the training
phase, the system collects sensor readings when the real user
moves her head with a music cue (following her pre-designed
movement pattern), and uses the collected data to build a
classifier. In the following discussion, we assume there is
only one real user for the device for the sake of simplicity.
An extension to support multiple users per device can be
realized through minor modifications, namely, by appropriately
indexing the users in the trained database.
In the online authentication phase, we collect sensor samples
during a user’s authentication attempt and label the samples
using the Headbanger classifier. The user is authenticated upon
a successful classification.
4
Cosine-distance with False Subjects
Cosine-distance with True Subject
2.0
Correlation-distance with False Subjects
Correlation-distance with True Subject
2.0
5000
1.6
1.4
1.2
1.8
4500
Distance Score
Distance Score
Distance Score
1.8
4000
1.6
1.4
3500
3000
1.2
1.0
0
DTW-distance with False Subjects
DTW-distance with True Subject
5500
5
10
15
Subject ID
20
25
30
(a) Cosine Distance
0
2500
5
10
15
Subject ID
20
25
30
(b) Correlation
0
5
10
15
Subject ID
20
25
30
(c) DTW
Fig. 3. (a) Cosine, (b) Correlation and (c) DTW distances are computed over 20 response time time series for each subject, with 28 subjects in total. For each
subject’s column, a red square represents the average distance among that subject’s own response times, while a blue dot represents the average distance to other
subjects’ response times. The results show that red distances are always lower than blue ones, suggesting that different subjects’ head-movement patterns are
distinguishable.
As illustrated in Figure 4, Headbanger involves the following key modules: sensor data collection and filtering, sample
distance computing, and classification. We will now discuss
these design aspects in more detail.
A. Sensor Data Collection and Filtering
The data collection step involves collecting motion sensor
data (we mainly focus on accelerometer in this study) while
the user makes head-movements in response to the music
cue with a duration of T seconds. The raw accelerometer
signals are collected at a sampling rate of r samples/sec. The
accelerometer data corresponding to one user, is a collection of
accelerometer readings on the 3D axis (x, y, and z) collected
over T -second duration, stored in a matrix with dimensionality
3 × rT . We will refer to this 3 × rT matrix as a sample. We
retain the duration T to be in the order of few seconds, as
frequency of human head movements are, intuitively, typically
in the order of few times per second.
Next, we filter the raw samples to remove noises due to
spurious movements such as vibration or shaking. We adopt a
low-pass digital Butterworth filter [4] and set a relaxed cut-off
frequency of 10Hz.
B. Sample Distance Computing
In this study, we build a distance-based classifier for its simplicity is well suited for wearable devices. There are various
ways of computing distances between two signals; we have
considered three popular distance-computing algorithms in this
study – Cosine (COS)distance, Correlation (COR)distance, and
dynamic-time warping (DTW)distance.
Suppose we have two time series Sa = (s1 , s2 , ..., sn ) and
Sb = (s1 , s2 , ..., sn ). Their COS distance is calculated as
Sa ·Sb
kSa k×kSb k ; The COR distance is calculated by dividing their
distance covariance by the product of their distance standard
deviations; The DTW distance is defined as the distance when
the optimal alignment between Sa and Sb has been determined
by ”warping” them non-linearly [2].
C. Classification
The classification step labels a test sample as “true” or
“false” depending upon whether its distance to the real user’s
training samples is below a threshold. Again, we choose this
method because it strikes a good balance between simplicity
and performance. Next, we explain how we build the classifier
and how to conduct online classification in detail:
1) Identifying Top-K Training Samples. Given M training
samples, we first identify the K samples that are closest
to all the training samples. For each training sample,
we calculate its average distance to the other M − 1
samples, and then choose those K samples that have
the lowest average distance values. These K samples are
empirical estimation of the centroid of the sample space,
and thus best represent the space among the collection of
the training samples. We refer to them as Top-K samples.
In our classifier, we focus on the Top-K training samples
instead of all the training samples because it does not only
incur much less computing overhead, but it also provides
much better robustness against noises in training data.
2) Establishing Distance Threshold. Suppose a sample, s, is
one of the Top-K samples. We have its distance scores
to the other M − 1 samples in the training set, from
which we can calculate the distance mean µs and distance
standard deviation σs . Then sample s’s distance threshold
is defined as (µs + nσs ), where n is referred to as the
threshold parameter for our distance-based classifier.
3) Classifying Test Sample. If we use a training sample s to
classify the test sample t, then t is labeled as a true sample
if the distance between s and t is less then s’s distance
threshold (µs + nσs ); otherwise, it is labelled as a false
sample. The strictness of this classifier is characterized
by the value of the threshold parameter, n; a large n can
increase the false acceptance rate while a small n value
can result in a high rejection rate of true samples.
4) Voting. We label the test sample according to all K Top-K
samples, and the final classification result is the majority
decision among all K individual classification results.
5
0.8
0.8
0.6
DTW-EER=4.43%
COS-EER=18.72%
COR-EER=24.94%
0.4
0.4
0.2
0.00.0
DTW
Cosine Distance
Correlation
0.2
0.4
0.6
False Accepted Rate
0.8
1.0
Fig. 5. The impact of the distance computing algorithm (i.e., DTW, cosine
distance, and Correlation). In this set of results, we varied the value of n from
0.0 to 10.0 with an increment of 0.1, resulting in 100 thresholds in each case.
We then plotted the TPR (y-axis) and FAR (x-axis) for each threshold. The
results show that DTW delivers much better accuracies than the other two
distance algorithms. As a result, in the remaining of the study, we will choose
DTW for distance computing.
Among the four steps outlined above, the first two steps belong
to the offline training phase, while the last two steps belong
to the online authentication phase. Finally, if the user’s test
sample is classified as “true” then the user is authenticated to
the device; otherwise, the user is rejected.
IV. E VALUATION
We conducted comprehensive evaluation of Headbanger through laboratory studies with human subjects – our
studies were approved by the Institutional Review Board (IRB)
of our institution. In the first phase of evaluation, we collected
from volunteer participants accelerometer sensor readings
with Google Glass. We analyzed these traces offline on a PC.
Our evaluation in this phase is primarily aimed at determining
the accuracy and robustness of Headbanger. In the second
phase of evaluation, we implemented a Headbanger app and
measured its processing latency. Our measurements suggest
that Headbanger is indeed light-weight and can be executed
on wearable devices such as Google Glass.
A. Authentication Accuracy of Headbanger
1) Participants: We had a total of 30 volunteer participants
for this experiment, including a total of 19 males and 11
females. The mean age of the participants was 29.7 years with
a standard deviation of 9.81 years. The youngest participant
was 23 years old while the eldest was 54 years old.
2) Procedure: Our first experiment setup aimed at emulating
the typical usage scenario of Headbanger for authentication,
where a user conducts head-movements in response to a music
cue played on the Google Glass device during a login attempt.
In this experiment, all participants were asked to wear a
Google Glass device. Participants who originally wore glasses
(e.g. corrective eyewear) were asked to remove their glasses
before conducting the experiment. The trials were conducted
0.2
0.00.0
1.00
0.98
0.96
0.94
K1-EER=4.43%
K3-EER=3.93%
0.92
0.90
0.000.020.040.060.080.10
K=1
K=3
0.2
0.4
0.6
False Accepted Rate
0.8
1.0
Fig. 6. The impact of different K values: K = 1 and K = 3. In this set of
results, we varied the value of n from 0.0 to 10.0 with an increment of 0.1,
resulting in 100 thresholds in each case. We then plotted the TPR (y-axis)
and FAR (x-axis) for each threshold. The results show that voting schemes
(K = 3) provide minor improvement on the performance. As a result, in the
remaining of the study, we will choose K = 1.
1.0
0.8
1.00
True Positive Rate
0.6
True Positive Rate
1.0
True Positive Rate
1.0
0.6
0.95
0.90
0.4
0.2
0.00.0
0.85
M30-EER=4.43%
M20-EER=6.22%
M10-EER=7.60%
0.80
0.00 0.05 0.10 0.15 0.20
0.2
0.4
0.6
False Accepted Rate
M=30
M=20
M=10
0.8
1.0
Fig. 7. The impact of training dataset size: 10, 20, and 30 samples. In this
set of results, we varied the value of n from 0.0 to 10.0 with an increment of
0.1, resulting in 100 thresholds in each case. We then plotted the TPR (y-axis)
and FAR (x-axis) for each threshold. The results show that having 30 samples
delivers the best performance without adding to the online authentication
computing overhead. As a result, in the remaining of the study, we will choose
to have 30 samples.
in an academic environment and overseen by one of our team
members. The Google Glass ran our data-collection app that
played a piece of music (music cue) for a specific duration, and
recorded the accelerometer sensor readings. The sensor readings were recorded into a text file that was stored in the Glass’s
memory and later transported to a PC for offline processing
through a Python script. The experiment was conducted in a
well-lit indoor academic laboratory environment.
During the course of a data collection session, the participants were allowed to take a break or withdraw from
data collection if they felt uncomfortable at any point. The
participants also could take a break for a minute after each
trial. Each trial lasted for the duration of the music cue played
6
10
EER (%)
8
6.14
6
4
6.65
4.43
2
0
10 Sec
6 Sec
5 Sec
Fig. 8. The EER value of Headbanger when users choose different music
cue lengths (10 sec, 6 sec and 5 sec). We have an EER value of 6.65% with a
5-second music cue, and 4.43% with a 10-second music cue, which is better
than results of many similar body-movement based authentication systems
(such as those in [13], [31]).
on the Glass, and a total of 40 such trials were conducted for
each of the 30 participants. The entire data collection effort
lasted over a duration of 60 days, of which 15 participants
conducted their trials in a single sitting over a period of two
hours, while the rest of the trails were spread over 3 days on
average per subject.
3) Metrics: We evaluate the accuracy of Headbanger using
metrics that are commonly used in evaluating authentication
systems, namely, true positive rate TPR (percentage of true
test samples that are correctly accepted), false acceptance rate
FAR (percentage of false test samples that are mistakenly
accepted), and true rejection rate TRR (percentage of true test
samples that are mistakenly rejected). These three metrics are,
however, largely dependent on the choice of the classification
threshold parameter in Headbanger– a strict threshold in the
classifier can lead to a high TRR value, while overly relaxing
the threshold can lead to a high FAR. Hence, in order to report
the threshold-independent performance, we also consider equal
error rate EER which is the percentage of errors when we have
F AR = T RR.
4) Tuning Important System Parameters: Figures 5-7
present the results on how the system’s performance is impacted by several important system parameters, namely, the
choice of sample distance computing algorithm, the number
of best representative training samples used for classification
(K), and the number of total training samples (M ). Based
upon these results, we tune the parameter values to balance
the tradeoff between authentication accuracy and data collection/computing overhead.
Recall that our classifier employs a distance threshold of
(µ + nσ), where µ and σ are calculated from the training
samples. For each parameter study, we varied the value of
n from 0.0 to 10.0 with an increment of 0.1, and had a
total of 100 threshold values. We then plotted the TPR on
y-axis and FAR on x-axis for each threshold value, resulting
curves referred to as Receiver Operating Characteristics (ROC)
curves.
Firstly, Figure 5 compares the performance of three distance
computing algorithms: COS (cosine), COR (correlation) and
DTW, assuming a single best representative training sample
K = 1, music cue duration of 10s, and 30 training samples
M = 30. Among these three algorithms, DTW fares much
better than the other two: its EER is 14.29% smaller than
that of COS distance, and 20.51% smaller than that of COR
distance. This is as expected because DTW is designed to
match the waveform of two signals [2] and thus outputs more
accurate distance score. As a result, in the remaining of this
study, we will use DTW for evaluation. Even though DTW
incurs more computation than the other two, our Google Glass
implementation shows that through software optimization, the
processing latency of DTW distance can be made very small
(see Section IV-C).
Secondly, Figure 6 compares the performance of two K
values: K=1 and K=3, assuming DTW distance, music cue duration of 10s, and 30 training samples. Recall that our classifier
compares the test sample against K best representative training
samples, generates K independent classification results, and
votes among them for the final authentication result. Hence,
we expect that considering top 3 samples will be better than
only considering the top 1 sample, as confirmed by the results
shown in Figure 6. However, we observe the improvement is
very marginal: the EER when K = 3 is only 0.5% smaller
than the EER when K = 1. On the other hand, having K = 3
incurs three times as much computation as having K = 1. As
a result, in the remaining of this study, we will use K = 1 for
evaluation.
Thirdly, Figure 7 compares the performance of three training
dataset sizes: M=10, 20, and 30 samples, assuming DTW
distance, K = 1, and music cue duration of 10s. We observe
that the EER when M = 30 is 1.79% smaller than the
EER when M = 20 and 3.17% smaller than that when
M = 10. We emphasize that the value of M has NO impact
on the computing overhead in the online authentication phase
because the classifier only compares the test sample against K
representative training samples. As a result, in the remaining
of this study, we will use M = 30 for the evaluation. Please
note that we don’t choose use a larger M value because the
benefit of having a larger M diminishes quickly after having
30 samples.
5) Authentication Accuracy Results: After tuning system
parameters to balance accuracy and computing overhead, we
next calculate the EER value of the resulting Headbanger system when users choose different music cue durations and
present the results in Figure 8. As soon as a user starts the
authentication procedure, how long the user must wait before
she receives the authentication result is an important qualityof-service metric, which we refer to as authentication latency.
Authentication latency consists of two parts: data input latency
and data processing latency, wherein the former is the time a
user spends on listening to the music cue and making head
movements while the latter is the time Headbanger spends on
computing the authentication result. Between these two parts,
data input latency is by far the bottleneck as the data processing
latency can be easily reduced (by better algorithms and/or
7
Fig. 9. Pictorial description of the nodding pattern employed by each target.
Target (a) only moved his head in the vertical direction, and his nodding
immediately follows each music beat (on-beat); target (b) only moved his
head in the vertical direction, but there is often a delay between his nodding
and the music beat (off-beat); target (c) occasionally combined shaking with
nodding, and on-beat. As a result, the nodding patterns in (b) and (c) are
slightly more complex than that in (a).
faster hardware) and/or hidden (by pipelining computing with
data collection). Unfortunately, data input latency is hard to
be reduced or hidden by the improvement in software or hardware. Recognizing that different users can tolerate different
latencies and desire different levels of authentication accuracy,
Headbanger allows the users to choose the music cue duration
(which has the same length as the data input latency).
From Figure 8, we observe that the EER value of Headbanger is 6.65% with a 5-second music cue, and 4.43% with
a 10-second music cue. We take the viewpoint that such error
rate is rather sufficient for personal head-worn devices that are
not used in hostile environments. Also, the data input latency
of 5-10 seconds is comparable with similar authentication
systems. For example, a gait-based authentication system [13]
delivers an EER of 7.3% after observing the user for 20 meters;
a pattern-swipe system [7] delivers at best a TPR of 98% and
a FAR of 50% with offline processing; an eye-blinking-based
authentication system [31] delivers the BAC=(TPR + FAR)/2
of 94.4%, with a processing latency of 32 seconds.
B. Authentication Robustness of Headbanger
After evaluating authentication accuracy for Headbanger,
we next study its robustness. In this study, we focus on
imitation attacks. For this purpose, we asked a number of
participants (attackers) to imitate simple nodding patterns from
three participants (targets) after watching the target’s video
for as long as they desire, and calculated their chances of
successful imitation - i.e., the attacker is mistakenly accepted
by Headbanger. We note that nodding is the simplest headmovement pattern and easiest to imitate; hence, the results
presented here represent the lower bound robustness that
Headbanger offers. In reality, users are more likely to employ
more sophisticated head-movement patterns, which will be
much harder to imitate.
1) Participants: We had a total of 37 volunteer participants,
including 31 males and 6 females. The average age of the
participants was 25.6 years with a standard deviation of 6.6
years. The youngest participant was 22 years old while the
eldest was 49 years old.
2) Procedure: Our second experiment aimed at emulating
a practical imitation attack scenario. In this experiment, three
targets recorded video when they were nodding with a music
cue. Note that the music is usually played via a bone conduction speaker or an earplug, and that it is difficult to use a
camcorder to capture the music sound in a noisy environment.
To address this concern, during recording, we set the speaker
volume to maximum and conducted the recording in a quiet
laboratory environment.
We divided the attackers into three groups, and asked each
group to imitated one target. In each session (consisting of 30
trials), the attacker could watch the video for as long as they
wish. Our system provided a feedback after each trial so that
the attackers could adjust their nodding pattern if they wanted.
After the attacker had 30 trials, we ended the session no matter
whether the attacker had succeeded or not. In each session, we
noted the total number of successes the attacker had as well
as the number of trails before the first successful imitation.
3) Results: Each of the three targets performed simple
nodding in this experiment; though simple, their nodding
patterns have varying complexity. As shown in Figure 9, Target
A moved his head vertically, each nodding on a music beat,
with no noticeable horizontal movement; target B also only
moved his head in the vertical direction, but there was always
a delay between his nodding and the corresponding music beat;
target C combined slight shaking along with nodding, and
closely followed music beats. Among these three targets, A
is the easiest, while the other two are slightly more complex.
Table I summarizes the results. Here, we use FAR to denote
the successful imitation rate because a successful imitation is
counted as a false acceptance in our system. The overall FAR
of the experiment is 6.94%, while the individual FARs for
the three targets are 15.83%, 2.77% and 2.72% accordingly.
Since target A had the easiest nodding pattern, 7 out of 12
participants could succeed at least once during their 30 trials,
while for targets B and C, the numbers are 3 out of 13 and 3
of 12, respectively. These results are very promising: when a
user employs slightly more complex head-movement patterns,
it becomes much harder for others to imitate (EER dropping
from more than 15% to around 2.7%).
Target
A
B
C
Overall
TABLE I.
No.
of
Attackers
12
13
12
38
No.
of
Successful Attackers
7
3
3
13
Average No. of
Trials Before
First Successful Login
10.33
14.33
17.67
13.17
FAR (%)
15.83
2.77
2.72
6.94
T HE ATTACK RESULTS SHOW THAT AS HEAD - MOVEMENT
PATTERNS BECOME MORE COMPLEX , IT BECOMES MUCH HARDER TO
IMITATE .
8
Fig. 10.
Software modules of Headbanger implementation
C. Headbanger Google Glass App Implementation
In the second phase of evaluation, we implemented Headbanger on Google glass as an authentication app. Figure 10
shows the main software modules in the app. Upon initiation
by the user, the app plays a music cue for a user-specified
duration. The user conducts head-movements in synchrony
with the music cue while the app records the accelerometer
data in parallel. At the end of the music cue duration the app
enters the data processing phase where the sensor readings are
input to the Headbanger’s software modules for processing.
Upon completion of data processing, the app responds with a
YES or NO textual output on the Google Glass screen. The
training phase is conducted offline on a PC, and we ensure that
the training set is readily available during the authentication
process.
1) Data Processing Latency: In Table II we report the
measured average processing latency of the Headbanger app
for music cue durations of 5, 6 and 10 seconds. The processing
latency is within 2 seconds for a 10-second data input, and is
less than 1 second (0.88s) for a 5-second input. We also report
the total latency breakdown for different software modules,
and find that DTW computation is the main bottleneck. In
our implementation, we used Fast DTW [33] which decreases
the computation complexity from O(n2 ) to O(n) without
compromising the authentication accuracy.
As we discussed earlier, the processing latency can be
further reduced, and it could be partially hidden if we pipeline
the data processing and data input. As a result, we believe
that our design of Headbanger is indeed light-weight and it
is realistic to run the Headbanger app on devices that have
comparable computing capabilities as the Google Glass.
V.
R ELATED W ORK
Several studies have looked at head or eye movements
for various purposes including user authentication. Harwin et
music cue
duration (s)
10
6
5
data processing latency (s)
1.93
1.15
0.88
time
Filtering
0.49
0.63
0.82
breakdown (%)
DTW
Thresholding
99.50
0.01
99.36
0.01
99.19
0.01
TABLE II.
M EASURED RESPONSE TIME OF Headbanger APP
IMPLEMENTATION ON G OOGLE G LASS WITH DIFFERENT MUSIC CUE
DURATIONS AND FOR K = 1. T HE RESPONSE TIME REPORTED HERE IS AN
AVERAGE OVER 20 TRIALS .
al. [17] are the first to use head gestures for human computer
interaction. Westeyn et al. [37] used eye-blinking pattern as
a unique feature for authentication. They achieved 82.02%
accuracy with 9 participants. Rogers et al. [31] proposed to
use unconscious blinking and head movement to identify a
user from a group of users. In this method, users were asked
to view rapidly changing pictures for 34 seconds before they
can be identified. Ishimaru et al. [21] comes close to our system
design; they proposed to combine the eye blinking frequency
from the infrared proximity sensor and head motion patterns
from accelerometer sensor on Google Glass to recognize
activities (e.g., reading, talking, watching TV, math problem
solving). We note that recognizing activities is usually much
easier than recognizing who is performing the activity, which
is our objective in this study.
There are also a number of physiological activity recognition
studies using computer vision methods [18], [26]. While [26]
primarily uses computer vision to detect head gestures, BioGlass [18] combines Google Glass’s accelerometers, gyroscope, and camera to extract physiological signals of the
wearer such as pulse and respiratory rates. Camera processing
on wearable devices, especially Google Glass is compute
intensive and has a high energy budget [27].
Accelerometers have long been used to sense, detect and
also recognize movements in other parts of the body; for
example, gait recognition requires sensing in areas such as
waist [1], pocket [14], arm [15], [29], leg [13] and ankle [12].
These techniques, though well known, may not be suitable
for on wearable devices due to complexity (computation and
energy) in the machine learning process.
Hand gesture and touchscreen dynamics are often coupled
for authenticating to a (touchscreen) device (see e.g. [5] for
a survey). A number of features [5], [32] (e.g. finger length,
hand size, swipe/zoom speed, acceleration, click gap, contact
size, pressure) and behavioral feature (e.g. touch location,
swipe/zoom length, swipe/zoom curvature, time, duration)
have been exploited as for authentication purposes [5], [9],
[11], [34].
VI.
C ONCLUDING R EMARKS AND F UTURE D IRECTION
As wearable devices are increasingly weaved into our everyday life, providing security to the data acquired by or accessed
through these devices becomes critically important. In this
study, we have developed a user authentication system that uses
head-movement patterns for direct authentication to a wearable
device. Compared to existing authentication solutions, the
proposed solution delivers accurate authentication, is robust
against imitation attacks, incurs low processing delays, and
offers great convenience to users.
Through an extensive evaluation that involves 95 users, we
observe that the average true acceptance rate of our approach
is at 95.57% and the false acceptance rates at 4.43%. We also
observe that even simple head-movement patterns only allow
less then 3% of the imitation attacks to succeed. We have
also implemented an app on Google Glass, and measured the
end-to-end processing delay of less than 2 seconds for a 10second data sample. As a result, we believe it is realistic for
9
the proposed authentication system to be executed on resourceconstrained devices such as smart-glass. We further believe
the proposed method can help enable wider deployment of
wearable devices and apps in our life. Towards this goal, in
our future work, we will focus on how to make the headmovement based user authentication approach more reliable
in real-world settings.
VII. ACKNOWLEDGMENTS
This material is based upon work supported by the National
Science Foundation under Grant Number 1228777. Any opinions, findings, and conclusions or recommendations expressed
in this material are those of the authors and do not necessarily
reflect the views of the National Science Foundation.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
H. J. Ailisto, M. Lindholm, J. Mantyjarvi, E. Vildjiounaite, and S.-M.
Makela. Identifying people from gait pattern with accelerometers. In
Defense and Security. International Society for Optics and Photonics,
2005.
D. J. Berndt and J. Clifford. Using dynamic time warping to find
patterns in time series. In ACM KDD AAAI workshop, 1994.
C. Bo, L. Zhang, X.-Y. Li, Q. Huang, and Y. Wang. Silentsense: silent
user identification via touch and movement behavioral biometrics. In
ACM MobiCom, 2013.
R. Challis and R. Kitney. The design of digital filters for biomedical
signal processing part 3: The design of butterworth and chebychev
filters. Journal of biomedical engineering, 1983.
G. D. Clark and J. Lindqvist. Engineering gesture-based authentication
systems. Pervasive Computing, IEEE, 14(1):18–25, 2015.
C. Cornelius, R. Peterson, J. Skinner, R. Halter, and D. Kotz. A wearable
system that knows who wears it. In ACM MobiSys, 2014.
A. De Luca, A. Hang, F. Brudy, C. Lindner, and H. Hussmann. Touch
me once and i know it’s you!: implicit authentication based on touch
screen patterns. In ACM CHI, 2012.
A. De Luca and J. Lindqvist. Is secure and usable smartphone
authentication asking too much? Computer, 48(5):64–68, May 2015.
T. Feng, J. Yang, Z. Yan, E. M. Tapia, and W. Shi. Tips: contextaware implicit user identification using touch screen in uncontrolled
environments. In ACM HotMobile, 2014.
Fitbit. http://en.wikipedia.org/wiki/Fitbit.
M. Frank, R. Biedert, E.-D. Ma, I. Martinovic, and D. Song. Touchalytics: On the applicability of touchscreen input as a behavioral biometric
for continuous authentication. Information Forensics and Security, IEEE
Transactions on, 8(1):136–148, 2013.
D. Gafurov, P. Bours, and E. Snekkenes. User authentication based on
foot motion. Signal, Image and Video Processing, 2011.
D. Gafurov, K. Helkala, and T. Søndrol. Biometric gait authentication
using accelerometer sensor. Journal of computers, 1(7):51–59, 2006.
D. Gafurov, E. Snekkenes, and P. Bours. Gait authentication and
identification using wearable accelerometer sensor. In IEEE AIAT, 2007.
D. Gafurov and E. Snekkkenes. Arm swing as a weak biometric for
unobtrusive user authentication. In IEEE IIHMSP, 2008.
Google glass. http://en.wikipedia.org/wiki/Google Glass.
W. Harwin and R. Jackson. Analysis of intentional head gestures
to assist computer access by physically disabled people. Journal of
biomedical engineering, 12(3):193–198, 1990.
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
J. Hernandez, Y. Li, J. M. Rehg, and R. W. Picard. Bioglass:
Physiological parameter estimation using a head-mounted wearable
device. In IEEE MobiHealth, 2014.
R. Hoyle, R. Templeman, D. Anthony, D. Crandall, and A. Kapadia.
Sensitive lifelogs: A privacy analysis of photos from wearable cameras.
In Proceedings of the 33rd Annual ACM Conference on Human Factors
in Computing Systems, pages 1645–1648. ACM, 2015.
R. Hoyle, R. Templeman, S. Armes, D. Anthony, D. Crandall, and
A. Kapadia. Privacy behaviors of lifeloggers using wearable cameras.
In ACM UbiComp, 2014.
S. Ishimaru, K. Kunze, K. Kise, J. Weppner, A. Dengel, P. Lukowicz,
and A. Bulling. In the blink of an eye: combining head motion and
eye blink frequency for activity recognition with google glass. In ACM
AH, 2014.
A. K. Jain, A. Ross, and S. Prabhakar. An introduction to biometric
recognition. Circuits and Systems for Video Technology, IEEE Transactions on, 14(1):4–20, 2004.
S. Jana, A. Narayanan, and V. Shmatikov. A scanner darkly: Protecting
user privacy from perceptual applications. In Security and Privacy (SP),
2013 IEEE Symposium on, pages 349–363. IEEE, 2013.
A. R. Jensenius. Action-sound: Developing methods and tools to study
music-related body movement. 2007.
Z. Jorgensen and T. Yu. On mouse dynamics as a behavioral biometric
for authentication. In ASIACCS, 2011.
R. Kjeldsen. Head gestures for computer control. In IEEE ICCV
Workshop, 2001.
R. LiKamWa, Z. Wang, A. Carroll, F. X. Lin, and L. Zhong. Draining
our glass: An energy and heat characterization of google glass. In
Proceedings of 5th Asia-Pacific Workshop on Systems, page 10. ACM,
2014.
F. Monrose and A. D. Rubin. Keystroke dynamics as a biometric for
authentication. Future Generation computer systems, 16(4):351–359,
2000.
F. Okumura, A. Kubota, Y. Hatori, K. Matsuo, M. Hashimoto, and
A. Koike. A study on biometric authentication based on arm sweep
action with acceleration sensor. In IEEE ISPACS, 2006.
T. Rahman, A. T. Adams, M. Zhang, E. Cherry, B. Zhou, H. Peng, and
T. Choudhury. Bodybeat: a mobile system for sensing non-speech body
sounds. In ACM MobiSys, 2014.
C. E. Rogers, A. W. Witt, A. D. Solomon, and K. K. Venkatasubramanian. An approach for user identification for head-mounted displays. In
Proceedings of the 2015 ACM International Symposium on Wearable
Computers, pages 143–146. ACM, 2015.
N. Sae-Bae, K. Ahmed, K. Isbister, and N. Memon. Biometric-rich
gestures: a novel approach to authentication on multi-touch devices. In
ACM CHI, 2012.
S. Salvador and P. Chan. Toward accurate dynamic time warping in
linear time and space. Intelligent Data Analysis, 2007.
M. Sherman, G. Clark, Y. Yang, S. Sugrim, A. Modig, J. Lindqvist,
A. Oulasvirta, and T. Roos. User-generated free-form gestures for
authentication: Security and memorability. In ACM MobiSys, 2014.
Smart watch. http://en.wikipedia.org/wiki/Smartwatch.
S. V. Stevenage, M. S. Nixon, and K. Vince. Visual analysis of gait as
a cue to identity. Applied cognitive psychology, 13(6):513–526, 1999.
T. Westeyn and T. Starner. Recognizing song-based blink patterns:
Applications for restricted and universal access. In IEEE FGR, 2004.
M. Zentner and T. Eerola. Rhythmic engagement with music in infancy.
Proceedings of the National Academy of Sciences, 2010.